zhangyue19921010 opened a new issue, #11331: URL: https://github.com/apache/paimon/issues/11331
### Search before asking - [x] I searched the issues and found no matching report. ### Paimon version 2.2-SNAPSHOT, master at `61f9981bc3b265da12446bf48eb0db441fb2da86`. ### Compute Engine Java API, JDK 11. ### Minimal reproduce step Use ORC with `orc.sarg.to.filter=true` and `orc.filter.use.selected=true`, without deletion vectors. Write an `id INT` column and a nested payload, then push a predicate on `id` while reading the payload. For example, write six rows with IDs 0 through 5 and `ROW(10 + id)` payloads, then read with `id >= 3`. The real-file regression returns IDs 3, 4, 5 but payloads 10, 10, 10. ARRAY and MAP payloads are also affected. An ordinary append table regression reproduces the problem through `TableRead.executeFilter()`. ### What doesn't meet your expectations? Nested values should stay aligned with their IDs. A selected ORC batch maps top-level positions to physical rows; nested adapters apply that selection again to physical row or child positions, returning incorrect values. Residual filtering cannot repair those values. ### Anything else? The proposed fix keeps row-group pruning for nested projections and lets the caller evaluate the residual predicate. Primitive projections can still use selected-row filtering. This gives up the optional row-level pushdown for nested projections until their vector consumers can handle it correctly. ### Are you willing to submit a PR? - [x] I'm willing to submit a PR! -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
